What is an AI agent, and what can it do for your business?

An AI agent is a program that takes a goal, decides on its own which steps to take and which tools to use to reach it, and checks the result of each step before moving on. Unlike an automation, nobody wrote out its path in advance.

In the US, "what is an AI agent" gets about 22,200 searches a month, according to OpenSEO (October 27, 2026), and the word now gets attached to almost any software with a language model inside. In June 2025, Gartner estimated that of the thousands of vendors selling agentic AI, only about 130 actually are, and coined "agent washing" for rebranding a chatbot, an assistant, or an RPA bot as an agent. This article separates what an agent is from what it isn't, explains how one works under the hood, and walks through five business tasks with the call we'd make on each: agent, automation, or template. It includes a real case where the right answer was not to use an agent, and the method we follow to start a project when one is warranted.

What is an AI agent?

An AI agent is a system in which a language model directs its own process: it picks a tool, uses it, looks at the result, and decides the next step until the task is done or it needs a person. The most useful definition we've found comes from Anthropic, which in December 2024 separated two things the market blurs. In workflows, the model and its tools follow code paths set in advance. With agents, the model controls how the task gets done. Google Cloud, on its AI agents page updated in 2026, describes them as software that pursues goals on a user's behalf, with the ability to reason, plan, and remember. Both definitions agree on the part that matters to a business: an agent acts on other systems, and it does so with some room to make its own decisions, room that someone has to define.

Inside, an agent has four parts. A language model that reasons about the task. Tools it acts with: reading an inbox, querying the CRM, searching internal documentation, writing to a spreadsheet, or opening a ticket. Context covering what it has already done and what it needs to know about the company. And a stopping condition that says when the task is finished, how many attempts it gets, and when it should stop and ask. Anthropic puts weight on that last part, recommending limits such as a maximum number of iterations and checkpoints where the agent waits for a person to respond. Without one, an agent that gets step two wrong keeps building on the mistake through steps three, four, and five, and the failure only surfaces at the end, after it has already touched several systems.

How is an AI agent different from a chatbot or an automation?

The difference is who decides the next step. With a chatbot, the person does: the program answers and waits for the next question. With a classic automation, such as a workflow that connects two apps or an RPA bot, whoever designed it decides, because every step is written down and runs the same way each time. With an AI workflow, the designer sets the path and the model handles certain steps, like classifying an email or summarizing a document. With an agent, the model chooses the path among the tools it's allowed to use. Moving across the table, you gain flexibility, lose predictability, and pay more per task. Anthropic is blunt about it: agentic systems trade latency and cost for better results, and you should start with the simplest solution that works, which often isn't an agent.

Google Cloud adds a distinction that clears up another common mix-up: the assistant. An AI assistant works alongside the user, responds to requests, and can recommend actions, but the user makes the call at each step. An agent receives a goal and acts proactively until it's met. A bot, in the same breakdown, follows preset rules and learns little. For a business, the practical question is how much decision-making it's willing to hand over on each task. An assistant that drafts a reply to a customer and waits for someone to send it carries little risk. An agent that sends that reply, updates the order in the ERP, and issues a refund needs narrow permissions, a log of every action, and a spending limit above which it asks first. The product label tells you none of this. What tells you is what it can touch, and whose sign-off it needs.

How does an AI agent work?

An agent works in a loop: it takes the goal, plans, uses a tool, observes the result, and decides whether to keep going, change its plan, ask someone, or stop. Anthropic describes it as a model using tools based on feedback from its environment, and stresses that at every step the agent needs ground truth it can check, such as a query result or a program's output, to know whether it's making progress. Say a customer asks about a missing order. The agent reads the email, looks up the order in the ERP, checks the carrier's tracking, sees the package is held at a warehouse, drafts a reply with the new delivery date, and, because the amount is over the set limit, prepares the refund for a person to approve. No diagram laid out those steps. The agent chose them as it went.

The loop also explains why an agent costs more than a workflow. Every pass is a call to the model, and a task a workflow handles in two calls can take an agent ten if it hesitates, repeats a search, or changes course. That's why agents fit best where the number of steps can't be predicted, and worst where the process is always the same. Anthropic suggests using them for open-ended problems, where it's hard or impossible to know in advance how many steps you'll need and you can't hard-code a path, and warns that their autonomy brings higher costs and compounding errors. The operational takeaway is that an agent gets designed with three decisions made before anyone writes a line of code: which tools it can use, with which permissions, and when it has to stop and hand the case to a person.

What can an AI agent do for a business?

An agent fits tasks with variable steps that cut across several systems, such as handling customer questions that require checking the order, the shipment, and the invoice. For tasks that always follow the same steps, an automation or an AI workflow does the job at lower cost and lower risk. The table covers five business tasks and the call we'd make on each. Three come from real conversations with AI buyers this year: a contract workflow one company already had running, the finance and purchasing area that same leadership team named as the one with the most manual hours, and a request to automate technical reports. The other two, customer service and lead qualification, are the uses our own AI consulting page offers. Only one of the five calls for a true agent.

The HR case shows the reasoning. The leadership team at a retail chain told us they already had a workflow running to add, terminate, or change employee contracts, linked to their outside payroll provider. That's a textbook task that doesn't need an agent: the steps never change, the data is personal, and a mistake has legal and payroll consequences. A rules-based automation does the job. Finance and purchasing work much the same way, with one difference. Invoices and purchase orders arrive in all sorts of formats, and reading them is real work for a language model, which pulls out the supplier, amounts, and dates. Entering that data into the ERP, on the other hand, runs on rules. That split between what the model reads and what the code does is what we mean by an AI workflow.

When don't you need an AI agent?

You don't need one when the process runs the same way every time, when part of the output has to be exact and verifiable, or when the real problem is something else. The clearest case we've seen is an accredited environmental inspection firm that writes ten to twelve technical soil reports a month, each one taking a technician four to seven days. That's 40 to 84 technician-days a month. They asked us to automate it with AI. When we compared three real reports, nearly half the body was word-for-word identical from one report to the next: methodology, confidentiality, sampling plan, and chain of custody. That part doesn't need a language model. It needs templates with conditional blocks filled in with each site's data, and it costs nothing per model call because there aren't any.

The compliance statement, which checks every measured value against the legal limits, runs on code and never on a generative model, because it's the part the firm's accreditation covers. AI handles the writing that sits on top of already-calculated data: the geological setting, the summary, and the commentary. The conclusions still belong to the technician who signs the report. The final design has three layers, and none of them is an agent. We share it because it's what we expect to find when a company asks for "an agent that writes the reports": most of the savings come from organizing the document and separating calculation from prose, and AI comes in afterward, on the part where it actually adds something. It's the diagnosis we'd run before quoting any agent at all.

How do you get started with AI agents in your company?

You start with an inventory of processes and systems, and choose the tool later. Before deciding whether a task calls for an agent, you need to know which systems the company already runs (its CRM, its ERP, its document management system) and what data lives in each, because an agent that doesn't connect to them stays a demo. Then you pick a single task and run a scoped pilot: one use case, working in real day-to-day operations, with business metrics set before you start, such as time per task, the share of cases the agent resolves without escalating, or errors caught in review. Whatever hits its metrics gets rolled out to more areas. Whatever doesn't gets fixed or shut down. That's how our AI consulting for businesses works, and its first step is telling you if your company isn't ready yet.

The order matters because agentic projects rarely fail because of the model. In the same June 2025 forecast, Gartner predicted that more than 40% of agentic AI projects will be canceled by the end of 2027, due to rising costs, unclear business value, or inadequate risk controls. Analyst Anushree Verma added that many use cases being pitched as agentic today don't need an agentic implementation. All three causes can be headed off during the diagnosis: cost, by choosing the simplest solution that works for each task; value, by setting the metric before the pilot starts; and risk, by putting in writing what the agent can touch and when it stops. In practice, you manage an agent the way you'd onboard a new hire: limited permissions and close supervision for the first few months.

What are the risks of an AI agent?

The main risk with an agent is that it acts: a chatbot's mistake is a bad answer, while an agent's mistake can be an email sent, an order changed, or a record deleted. Anthropic warns that agents' autonomy brings higher costs and errors that compound from one step to the next, and recommends extensive testing in sandboxed environments, with guardrails, before going live. In practice that comes down to four controls. Least privilege: the agent can only reach the systems and actions its task requires. Logging: every action is recorded with its reason, so it can be audited. Limits: amounts, volumes, or case types above which the agent prepares the action and waits for a person to approve it. And data: which customer or employee information it can read, where it's processed, and who's accountable for it.

Google Cloud also lists the areas where agents fall short: tasks that call for empathy or reading complex social situations, decisions with high ethical stakes, such as a medical diagnosis or a court ruling, and applications where the cost of building and running one doesn't pay off for a small organization. The list works well as a first filter. If a task lands in one of those areas, the agent can prepare the work by gathering data, drafting, or summarizing a file, but a person still makes the decision. Everywhere else, almost anything can be done with an agent. The useful question is whether the result justifies the cost per task and the work of supervising it, compared with an automation that would do the same job with rules.

Coming up in this series

This article kicks off a series on AI in business. The next pieces go deeper on each decision:

• AI process automation: what it looks like in each department. Coming soon.
• ChatGPT for business: what works, what doesn't, and how to roll it out. Coming soon.
• AI for business: a decision guide for executives. Coming soon.
• AI in marketing: case studies with numbers. Coming soon.
• AI consulting: what it includes and what it should cost. Coming soon.

For the other side of the story, how AI agents read websites, see llms.txt: what it is and what it's for, how to check whether ChatGPT cites your site, and what GEO is.

FAQ

Does an AI agent make sense for your company?

We start with a diagnosis of your processes and systems, and tell you which tasks call for an agent, which are better handled with an automation or a template, and which are best left alone for now. See our AI consulting.